AutoPHD by LanternW

An open-source, local-first harness for long-horizon research agents.
Research-career, project, and stage-task loops with mentor gates, evidence tracking, periodic reflection, and a local Web UI.

English · 简体中文 · Documentation

AutoPHD helps researchers and agent developers run inspectable, stateful research workflows instead of asking a model for a single final report. It keeps task state and artifacts in local workspaces while using configured model providers or external agent backends for inference.

At a glance

Question	Answer
What is it?	A research-agent harness that maintains projects, stages, milestones, evidence, meetings, and reusable task memory.
Who is it for?	Researchers and developers evaluating long-running, tool-using research agents in a local workspace.
How is it different from a deep-research agent?	It emphasizes an ongoing research process and review gates, not one query followed by one synthesized report.
What can I see in 30 seconds?	Start the Web UI and inspect profile, task, milestone, activity, log, runtime, and conversation views.
Is it production-ready?	Not recommended. Any work under consideration for publication should undergo peer review.

See it locally

Python 3.10 or later is required.

python -m pip install -r requirements.txt
python WebUI/server.py

Open http://127.0.0.1:8765. The server prints a local URL similar to:

AutoPHD WebUI local: http://127.0.0.1:8765

You can inspect the interface and create a research profile without an API call. Starting an agent task requires a configured model provider or supported external backend and may incur provider costs.

Why AutoPHD?

Many research assistants optimize for a useful answer to one prompt. Long-running research has different failure modes: objectives drift, weak evidence survives into later drafts, failed attempts disappear, and a polished manuscript can hide an unreliable process. AutoPHD makes the process itself inspectable through persistent state, staged work, mentor feedback, evidence records, and periodic review.

Core capabilities

• Three-layer workflow: research-career goals, research projects, and stage-specific tasks are represented as linked state rather than a single conversation.
• Mentor-gated milestones: mentor meetings and milestone evaluation create explicit checkpoints and feedback loops.
• Evidence-aware research: tools record research evidence, audit integrity, and check stage readiness.
• Durable local artifacts: profiles, logs, knowledge, papers, milestones, and project files remain inspectable in local workspaces.
• Local Web UI: create profiles and tasks, then follow activity, milestones, runtime signals, logs, and conversations.
• Extensible agent runtime: the harness supports its OpenAI-compatible path and configurable external agent backends; exact provider behavior depends on local configuration.

How it works

1. A profile defines a research mission, graduation goal, model settings, and mentor cadence.
2. AutoPHD creates a task workspace under RunningSpace/ and maintains structured pipeline and milestone state.
3. The research agent uses skills and tools to plan, collect evidence, manage artifacts, and advance work.
4. Mentor meetings and stage gates challenge progress and can require more evidence before advancement.
5. Logs, reflections, and reusable knowledge preserve an inspectable record for later review.

See the architecture documentation for component and data-flow details.

Five-minute quick start

1. Start the Web UI with the commands above.
2. Enter an OpenAI-compatible API base URL and API key in Configuration, then save it locally.
3. Create a research profile with a bounded research direction, measurable goal, model, and mentor settings.
4. Open Tasks, create a task from the profile, and start it.
5. Follow the activity, milestones, logs, runtime signals, and conversations from the task page.

Profiles are stored in configs/profiles/. Each task receives a copied harness and research workspace under RunningSpace/. Do not commit credentials or private research data.

Example input and expected output

An initial profile might define:

Research mission: Map robust evaluation methods for evidence-grounded literature agents.
Graduation goal: Produce a reproducible benchmark protocol and one evaluated baseline.
First milestone: Define inclusion criteria, failure categories, and an auditable test set.

Expected artifacts are process records, not a guaranteed scientific result:

RunningSpace/<task-id>/
├── status/                       # pipeline and current milestone state
├── data/<research-project>/      # research artifacts and evidence
├── .LOG/                         # execution and meeting traces
└── .HUMAN_LOG/                   # human-readable progress records

Exact files evolve with the selected workflow and tools. A successful run still requires a person to validate sources, methods, code, statistics, and conclusions.

Appropriate and inappropriate uses

Good evaluation targets include long-horizon agent orchestration, milestone and stage-gate design, research trace inspection, evidence-management workflows, and human/agent supervision experiments.

Do not treat AutoPHD as an autonomous author of trustworthy papers, a replacement for domain experts or ethics review, a hardened execution sandbox, or proof that an output is novel or reproducible.

Research integrity

AutoPHD includes evidence records, integrity audits, mentor review, and stage gates, but these mechanisms reduce risk rather than establish truth. Users remain responsible for source verification, permissions, privacy, experimental controls, statistical validity, authorship disclosure, and compliance with institutional and venue policies.

Read the full research-integrity policy.

Roadmap

• Publish a versioned end-to-end demonstration with a complete transcript.
• Release reproducible benchmark tasks with time, cost, model, environment, and failure reporting.
• Publish reviewed case studies tied to exact commits.
• Expand automated tests around stage gates, evidence provenance, and unsafe runtime boundaries.
• Tag the first semantic release after the public interface and installation path stabilize.

Contributing, security, and citation

• Contribution workflow: CONTRIBUTING.md
• Security reporting and deployment boundaries: SECURITY.md
• Release history: CHANGELOG.md
• Machine-readable citation metadata: CITATION.cff

When reporting a result, cite the exact release or commit, model/provider, configuration, input data, elapsed time, cost, full trace, and human-review outcome. This repository is licensed under the GNU AGPL v3.0.